Directed Diffusion A Scalable and Robust Communication Paradigm

Directed Diffusion: A Scalable and Robust Communication Paradigm for Sensor Networks Chalermek Intanagonwiwat, Ramesh Govindan and Deborah Estrin (Mobi. COM '00)

Distributed Microsensing l Characteristics of future sensor nodes may include: – – – l Matchbox size Battery power Power-conserving processors clocked at several hundred Mhz Program and data memory of several tens of Mbytes Radio modem Energy efficient MAC layer Such hardware will enable distributed microsensing, in which a collection of nodes coordinate to achieve a larger sensing task

Possible Scenarios l Sensor networks may be deployed in – Inhospitable physical environments l l – Remote geographic regions Toxic urban locations Difficult to access environments l l Large industrial plants Aircraft interiors

Critical Design Constraints l l l Power conservation Scalability—must scale to several thousands of sensor nodes Robustness—high frequency of node failure expected

Directed Diffusion l Directed diffusion is a new data dissemination paradigm for sensor networks – – Data-centric Determined by localized interactions: hop-to-hop rather than end-to-end l l – Long-range communication requires more energy Hop-to-hop communication provides link diversity, helps overcome obstacles Incorporates application-specific semantics

Directed Diffusion Elements l Named data – l Described with attributevalue pairs Interests – l Describe sensing tasks Gradients – l Network paths Reinforcement – Manages gradients

Naming l Task descriptions are named by a list of attribute-value pairs: Intuitively, the task description specifies an interest for data matching the attributes. For this reason, such a task description is called an interest

Interest Dissemination C C’s Interest cache Interests Gradients source B Sink disseminates interest for a four-legged animal (~36 bytes). Initial interval is large. sink

Interest Dissemination C C’s Interest cache Interests Gradients source B sink Every node contains an interest cache, with separate entries for distinct interests. Entries do not contain info about sinks and therefore scale well. Overlapping entries may be aggregated for efficiency. Interests must be periodically refreshed by sink.

Interest Dissemination C source C’s Interest cache Interests Gradients Sink: 1 s | B: 1 s source: 1 s B Each interest cache entry contains a list of gradients; events that match interest entries are propagated back to the sink via these gradients. Gradient entries contain locally unique neighbor IDs, data rates, and interval (not shown) attributes. sink In the absence of information about which sensor nodes are likely to be able to satisfy an interest, interests are broadcasted to all neighbors. However, a node may suppress a received interest if it recently re-sent a matching request.

C’s Data cache Data Propagation C 1 eps EVENT C’s Interest cache Interests 1 eps source 1 eps B Gradients Sink: 1 s | B: 1 s source: 1 s 1 eps Initial interests request data at slow rates (e. g. 1 event per second). sink A sensor node that detects a target searches its interest cache for a matching entry; if it finds one, it begins sending data messages (~64 bytes) towards the sink via its gradient list at the highest specified rate.

C’s Data cache Data Propagation C 1 eps EVENT C’s Interest cache match Interests 1 eps source 1 eps B Gradients Sink: 1 s | B: 1 s source: 1 s 1 eps Upon receiving a data message, nodes check their interest caches. If no match is found, the data message is silently dropped. sink If a match is found, the node checks its data cache, which keeps track of recently seen data items. If no data cache entry matches the message, a new entry is made in the data cache and the message is re-sent to the node’s neighbors. If a data cache entry matches the data message, the message is silently dropped.

C’s Data cache Reinforcement C 100 1 eps EVENT C’s Interest cache Interests 1 eps source 1 eps B Gradients Sink: . 01 s S: 1 s | B: 1 s source: 1 s 1 eps 100 eps 1 eps After the sink starts receiving these low data rate events, it sink reinforces one particular neighbor in order to “draw down” higher quality (higher data rate) events. It does this explicitly by re-sending the original interest message, but with a smaller interval value, to the empirically low delay path node. Nodes update their caches and can then propagate reinforcement messages according to local policies. For example, the node might choose that neighbor from whom it first received the latest event matching the interest

Considerations l l Embedding application semantics in communication logic allows for optimizations such as loop prevention and downconversion (for instance, interpolating high rate messages for a low rate receiver) Negative reinforcement is used to prune superfluous gradients

Negative Reinforcement l l Could use time outs or explicit degrade messages as negative reinforcement mechanisms Orthogonal to the mechanism, NR controls can be propagated according to a number of different rules – E. g. : negatively reinforce that neighbor from which no new events have been received within a window of N events or T time units

Network Topology l This paradigm works with multiple sources (but sinks may draw redundant data) and multiple sinks hosting identical interests (in which case the second sink can immediately draw down high quality via its cache)

Local Repair l Reinforcement rules can be applied by intermediate nodes to repair faulty links: – – Node C can discover better path by requesting higher rates from non-faulty neighbors Reinforcement must be applied carefully to prevent all downstream nodes from doing the same, which will result in discovery of a good path, but will waste resources

Design Parameters

Evaluation: Metrics l Average Dissipated Energy – – l Average Delay – l Measures the ratio of total dissipated energy per node in the network to the number of distinct events seen by sinks Computes average work done by a node as well as the overall lifetime of sensor nodes Measures the average one-way latency between transmitting an event and receiving it at a sink Event Delivery Ratio – Ratio of the number of distinct events received to the number originally sent

Evaluation: Setup l l l ns-2 simulated 1. 6 Mbps 802. 11 MAC layer (with RTS/CTS—not optimal for this application because nodes expend as much power idle as when receiving) 5 different sensor fields, ranging from 50 nodes (at 160 m x 160 m) to 250 nodes (at same average density) DD compared against Flooding and Omniscient Multicast (which approximately represents achievable IP-based performance)

Evaluation: Comparative Results DD dissipates less energy than OM because of innetwork aggregation, in which intermediate nodes suppress duplicate location estimates. However, the savings is not equal to the number of identical sources because 1)both schemes spend significant energy listening, and 2)design parameters are conservative and frequently draw down high quality data along multiple paths.

Evaluation: Comparative Results Flood’s poor performance (which is exclusively dependent upon broadcast) an artifact of the MAC layer: random delay imposed to avoid broadcast collisions. MAC with TDMA might equalize Flood’s performance.

Evaluation: Dynamics l l l Create node failures in the paths diffusion is most likely to use, as well as creating random failures elsewhere in the network At any instant, 10 or 20% of nodes unusable No “settling time” between failures

Evaluation: Dynamics Performance actually improves in some cases because overly-conservative reinforcement rules maintain several high-quality paths during normal operation, some of which fail due to node failures in this experiment, thus decreasing energy consumption.

Evaluation: Dynamics

Evaluation: Dynamics

Evaluation: Other Factors Even conservative negative reinforcement prunes enough redundant paths to achieve substantial energy savings

Evaluation: Other Factors DD expends substantially less energy when it can aggregate (or suppress) duplicate messages. Conservative negative reinforcement rules account for decrease in energy consumption without diffusion in larger networks: in such networks, the lengths of paths pruned by negative reinforcement are sizeable and result in proportionally sizeable decreases in energy consumption.

Evaluation: Other Factors If there are no gains to be had from wise transmission management (i. e. if idle mode consumes as much power as receive mode), then there is no point in implementing complicated communication schemes because idle time will dominate the performance of all schemes.

Related Work l Distributed Sensor Networks – l l l Reaction-diffusion models for morphogenesis and models of ant colony behavior Ad hoc unicast routing (particularly reactive protocols) Multicast routing protocols – – l l WINS and Piconet reinforcements similar to join/prune interest dissemination and gradient setup similar to datadriven shortest-path tree setup Router assist for localized error recovery in reliable multicast Web caching

Conclusions l l l Diffusion is data-centric All communication is neighbor-to-neighbor, not endto-end No routers—each node can interpret all messages No globally unique IDs (but locally unique IDs needed) Application-specific semantics embedded in communication http: //www. isi. edu/scadds/projects/diffusion. html